search

penincillin / DREAM

This is the public repository for our accepted CVPR 2018 paper "Pose-Robust Face Recognition via Deep Residual Equivariant Mapping"
http://mmlab.ie.cuhk.edu.hk/projects/DREAM/
BSD 2-Clause "Simplified" License
389 stars 98 forks source link

使用了DREAM模型后，为什么两个不同的人脸相似度反而提升了 #16

Closed songyy137222 closed 6 years ago

songyy137222 commented 6 years ago

两个相同人脸相似度是提升的，但提升特别小，我是用你提供的直接数据训练的DREAM模型（switching）

penincillin commented 6 years ago

Thanks for asking, I get confused by question, you use the provided data or provide pretrained model? What's more, English is prefered :)

songyy137222 commented 6 years ago

I'm sorry.I get the DREAM model with the provided data. (python branch_train.py) The test data is 256 dimensional vector with mobilefacenet.

songyy137222 commented 6 years ago

@penincillin

penincillin commented 6 years ago

The data we provided is too little to train a useful model. The only purpose of providing this data is to make sure you can go through the whole pipeline of our code and have a idea of the correct format of the data. To train a useful model, you should use much more data than sample data provided by us.

songyy137222 commented 6 years ago

@penincillin Iget it,thanks!!!

ysc703 commented 5 years ago

Hi, @songyy137222 Have you got a good result by using branch_train? I trained the DREAM branch with my own dataset(58k images of 6k IDs), but the loss didn't decrease.

Thanks!

jolinlinlin commented 5 years ago

@songyy137222 hello, Can you please tell me how to get 21-point of face to alignment?